diluvian

Flood filling networks for segmenting electron microscopy of neural tissue.

PyPI Release
Documentation
License
Build Status

Diluvian is an implementation and extension of the flood-filling network (FFN) algorithm first described in [Januszewski2016]. Flood-filling works by starting at a seed location known to lie inside a region of interest, using a convolutional network to predict the extent of the region within a small field of view around that seed location, and queuing up new field of view locations along the boundary of the current field of view that are confidently inside the region. This process is repeated until the region has been fully explored.

As of December 2017 the original paper's authors have released their implementation.

Quick Start

This assumes you already have CUDA installed and have created a fresh virtualenv. See installation documentation for detailed instructions.

Install diluvian and its dependencies into your virtualenv:

pip install diluvian

For compatibility diluvian only requires TensorFlow CPU by default, but you will want to use TensorFlow GPU if you have installed CUDA:

pip install 'tensorflow-gpu==1.3.0'

To test that everything works train diluvian on three volumes from the CREMI challenge:

diluvian train

This will automatically download the CREMI datasets to your Keras cache. Only two epochs will run with a small sample set, so the trained model is not useful but will verify Tensorflow is working correctly.

To train for longer, generate a diluvian config file:

diluvian check-config > myconfig.toml

Now edit settings in the [training] section of myconfig.toml to your liking and begin the training again:

diluvian train -c myconfig.toml

For detailed command line instructions and usage from Python, see the usage documentation.

Limitations, Differences, and Caveats

Diluvian may differ from the original FFN algorithm or make implementation choices in ways pertinent to your use:

• By default diluvian uses a U-Net architecture rather than stacked convolution modules with skip links. The authors of the original FFN paper also now use both architectures (personal communication). To use a different architecture, change the factory setting in the [network] section of your config file.
• Rather than resampling training data based on the filling fraction f_a, sample loss is (optionally) weighted based on the filling fraction.
• A FOV center's priority in the move queue is determined by the checking plane mask probability of the first move to queue it, rather than the highest mask probability with which it is added to the queue.
• Currently only processing of each FOV is done on the GPU, with movement being processed on the CPU and requiring copying of FOV data to host and back for each move.

[Januszewski2016]

Michał Januszewski, Jeremy Maitin-Shepard, Peter Li, Jorgen Kornfeld, Winfried Denk, and Viren Jain. Flood-filling networks. arXiv preprint arXiv:1611.00421, 2016.